CVPR 2024 Papers

Spurious correlations can cause strong biases in deep neural networks impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels training a debiased model from a limited amount of both annotations is still an open question. To address this issue we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically we first pretrain a biased encoder in a self-supervised manner with the rank regularization serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task serving as a boosting to effectively debias the main model. Remarkably the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and in some cases even outperforms state-of-the-art supervised debiasing approaches.

In this paper we delve into the nuanced challenge of tailoring the Segment Anything Models (SAMs) for integration with event data with the overarching objective of attaining robust and universal object segmentation within the event-centric domain. One pivotal issue at the heart of this endeavor is the precise alignment and calibration of embeddings derived from event-centric data such that they harmoniously coincide with those originating from RGB imagery. Capitalizing on the vast repositories of datasets with paired events and RGB images our proposition is to harness and extrapolate the profound knowledge encapsulated within the pre-trained SAM framework. As a cornerstone to achieving this we introduce a multi-scale feature distillation methodology. This methodology rigorously optimizes the alignment of token embeddings originating from event data with their RGB image counterparts thereby preserving and enhancing the robustness of the overall architecture. Considering the distinct significance that token embeddings from intermediate layers hold for higher-level embeddings our strategy is centered on accurately calibrating the pivotal token embeddings. This targeted calibration is aimed at effectively managing the discrepancies in high-level embeddings originating from both the event and image domains. Extensive experiments on different datasets demonstrate the effectiveness of the proposed distillation method. Code in https://github.com/happychenpipi/EventSAM.

Segment Anything Model (SAM) has emerged as a transformative approach in image segmentation acclaimed for its robust zero-shot segmentation capabilities and flexible prompting system. Nonetheless its performance is challenged by images with degraded quality. Addressing this limitation we propose the Robust Segment Anything Model (RobustSAM) which enhances SAM's performance on low-quality images while preserving its promptability and zero-shot generalization. Our method leverages the pre-trained SAM model with only marginal parameter increments and computational requirements. The additional parameters of RobustSAM can be optimized within 30 hours on eight GPUs demonstrating its feasibility and practicality for typical research laboratories. We also introduce the Robust-Seg dataset a collection of 688K image-mask pairs with different degradations designed to train and evaluate our model optimally. Extensive experiments across various segmentation tasks and datasets confirm RobustSAM's superior performance especially under zero-shot conditions underscoring its potential for extensive real-world application. Additionally our method has been shown to effectively improve the performance of SAM-based downstream tasks such as single image dehazing and deblurring.

Recent studies have drawn attention to the untapped potential of the "star operation" (element-wise multiplication) in network design. While intuitive explanations abound the foundational rationale behind its application remains largely unexplored. Our study attempts to reveal the star operation's ability of mapping inputs into high-dimensional non-linear feature spaces--akin to kernel tricks--without widening the network. We further introduce StarNet a simple yet powerful prototype demonstrating impressive performance and low latency under compact network structure and efficient budget. Like stars in the sky the star operation appears unremarkable but holds a vast universe of potential. Our work encourages further exploration across tasks with codes available at https://github.com/ma-xu/Rewrite-the-Stars.

The machine learning community has witnessed a drastic change in the training pipeline pivoted by those "foundation models" with unprecedented scales. However the field of adversarial training is lagging behind predominantly centered around small model sizes like ResNet-50 and tiny and low-resolution datasets like CIFAR-10. To bridge this transformation gap this paper provides a modern re-examination with adversarial training investigating its potential benefits when applied at scale. Additionally we introduce an efficient and effective training strategy to enable adversarial training with giant models and web-scale data at an affordable computing cost. We denote this newly introduced framework as AdvXL. Empirical results demonstrate that AdvXL establishes new state-of-the-art robust accuracy records under AutoAttack on ImageNet-1K. For example by training on DataComp-1B dataset our AdvXL empowers a vanilla ViT-g model to substantially surpass the previous records of l_ infinity - l_ 2 - and l_ 1 -robust accuracy by margins of 11.4% 14.2% and 12.9% respectively. This achievement posits AdvXL as a pioneering approach charting a new trajectory for the efficient training of robust visual representations at significantly larger scales. Our code is available at https://github.com/UCSC-VLAA/AdvXL.

The goal of interactive image segmentation is to delineate specific regions within an image via visual or language prompts. Low-latency and high-quality interactive segmentation with diverse prompts remain challenging for existing specialist and generalist models. Specialist models with their limited prompts and task-specific designs experience high latency because the image must be recomputed every time the prompt is updated due to the joint encoding of image and visual prompts. Generalist models exemplified by the Segment Anything Model (SAM) have recently excelled in prompt diversity and efficiency lifting image segmentation to the foundation model era. However for high-quality segmentations SAM still lags behind state-of-the-art specialist models despite SAM being trained with x100 more segmentation masks. In this work we delve deep into the architectural differences between the two types of models. We observe that dense representation and fusion of visual prompts are the key design choices contributing to the high segmentation quality of specialist models. In light of this we reintroduce this dense design into the generalist models to facilitate the development of generalist models with high segmentation quality. To densely represent diverse visual prompts we propose to use a dense map to capture five types: clicks boxes polygons scribbles and masks. Thus we propose SegNext a next-generation interactive segmentation approach offering low latency high quality and diverse prompt support. Our method outperforms current state-of-the-art methods on HQSeg-44K and DAVIS quantitatively and qualitatively.

Developing generalizable manipulation skills is a core challenge in embodied AI. This includes generalization across diverse task configurations encompassing variations in object shape density friction coefficient and external disturbances such as forces applied to the robot. Rapid Motor Adaptation (RMA) offers a promising solution to this challenge. It posits that essential hidden variables influencing an agent's task performance such as object mass and shape can be effectively inferred from the agent's action and proprioceptive history. Drawing inspiration from RMA in locomotion and in-hand rotation we use depth perception to develop agents tailored for rapid motor adaptation in a variety of manipulation tasks. We evaluated our agents on four challenging tasks from the Maniskill2 benchmark namely pick-and-place operations with hundreds of objects from the YCB and EGAD datasets peg insertion with precise position and orientation and operating a variety of faucets and handles with customized environment variations. Empirical results demonstrate that our agents surpass state-of-the-art methods like automatic domain randomization and vision-based policies obtaining better generalization performance and sample efficiency.

Recent advancements in language models pre-trained on large-scale corpora have significantly propelled developments in the NLP domain and advanced progress in multimodal tasks. In this paper we propose a Parameter-Efficient multimodal language model learning strategy named QaP (Querying as Prompt). Its core innovation is a novel modality-bridging method that allows a set of modality-specific queries to be input as soft prompts into a frozen pre-trained language model. Specifically we introduce an efficient Text-Conditioned Resampler that is easy to incorporate into the language models which enables adaptive injection of text-related multimodal information at different levels of the model through query learning. This approach effectively bridges multimodal information to the language models while fully leveraging its token fusion and representation potential. We validated our method across four datasets in three distinct multimodal tasks. The results demonstrate that our QaP multimodal language model achieves state-of-the-art performance in various tasks with training only 4.6% parameters.

Multi-modality large language models (MLLMs) as represented by GPT-4V have introduced a paradigm shift for visual perception and understanding tasks that a variety of abilities can be achieved within one foundation model. While current MLLMs demonstrate primary low-level visual abilities from the identification of low-level visual attributes (e.g. clarity brightness) to the evaluation on image quality there's still an imperative to further improve the accuracy of MLLMs to substantially alleviate human burdens. To address this we collect the first dataset consisting of human natural language feedback on low-level vision. Each feedback offers a comprehensive description of an image's low-level visual attributes culminating in an overall quality assessment. The constructed Q-Pathway dataset includes 58K detailed human feedbacks on 18973 multi-sourced images with diverse low-level appearance. To ensure MLLMs can adeptly handle diverse queries we further propose a GPT-participated transformation to convert these feedbacks into a rich set of 200K instruction-response pairs termed Q-Instruct. Experimental results indicate that the Q-Instruct consistently elevates various low-level visual capabilities across multiple base models. We anticipate that our datasets can pave the way for a future that foundation models can assist humans on low-level visual tasks.

Low-precision quantization is recognized for its efficacy in neural network optimization. Our analysis reveals that non-quantized elementwise operations which are prevalent in layers such as parameterized activation functions batch normalization and quantization scaling dominate the inference cost of low-precision models. These non-quantized elementwise operations are commonly overlooked in SOTA efficiency metrics such as Arithmetic Computation Effort (ACE). In this paper we propose ACEv2 - an extended version of ACE which offers a better alignment with the inference cost of quantized models and their energy consumption on ML hardware. Moreover we introduce PikeLPN a model that addresses these efficiency issues by applying quantization to both elementwise operations and multiply-accumulate operations. In particular we present a novel quantization technique for batch normalization layers named QuantNorm which allows for quantizing the batch normalization parameters without compromising the model performance. Additionally we propose applying Double Quantization where the quantization scaling parameters are quantized. Furthermore we recognize and resolve the issue of distribution mismatch in Separable Convolution layers by introducing Distribution-Heterogeneous Quantization which enables quantizing them to low-precision. PikeLPN achieves Pareto-optimality in efficiency-accuracy trade-off with up to 3X efficiency improvement compared to SOTA low-precision models.

Personalized Federated Learning (pFL) has emerged as a promising solution to tackle data heterogeneity across clients in FL. However existing pFL methods either (1) introduce high computation and communication costs or (2) overfit to local data which can be limited in scope and vulnerable to evolved test samples with natural distribution shifts. In this paper we propose PerAda a parameter-efficient pFL framework that reduces communication and computational costs and exhibits superior generalization performance especially under test-time distribution shifts. PerAda reduces the costs by leveraging the power of pretrained models and only updates and communicates a small number of additional parameters from adapters. PerAda achieves high generalization by regularizing each client's personalized adapter with a global adapter while the global adapter uses knowledge distillation to aggregate generalized information from all clients. Theoretically we provide generalization bounds of PerAda and we prove its convergence to stationary points under non-convex settings. Empirically PerAda demonstrates higher personalized performance (+4.85% on CheXpert) and enables better out-of-distribution generalization (+5.23% on CIFAR-10-C) on different datasets across natural and medical domains compared with baselines while only updating 12.6% of parameters per model. Our code is available at https://github.com/NVlabs/PerAda.

Recently some large kernel convnets strike back with appealing performance and efficiency. However given the square complexity of convolution scaling up kernels can bring about an enormous amount of parameters and the proliferated parameters can induce severe optimization problem. Due to these issues current CNNs compromise to scale up to 51x51 in the form of stripe convolution (i.e. 51x5+5x51) and start to saturate as the kernel size continues growing. In this paper we delve into addressing these vital issues and explore whether we can continue scaling up kernels for more performance gains. Inspired by human vision we propose a human-like peripheral convolution that efficiently reduces over 90% parameter count of dense grid convolution through parameter sharing and manage to scale up kernel size to extremely large. Our peripheral convolution behaves highly similar to human reducing the complexity of convolution from O(K^2) to O(logK) without backfiring performance. Built on this we propose Parameter-efficient Large Kernel Network (PeLK). Our PeLK outperforms modern vision Transformers and ConvNet architectures like Swin ConvNeXt RepLKNet and SLaK on various vision tasks including ImageNet classification semantic segmentation on ADE20K and object detection on MS COCO. For the first time we successfully scale up the kernel size of CNNs to an unprecedented 101x101 and demonstrate consistent improvements.

Diffusion MRI (dMRI) non-invasively maps brain white matter yet necessitates denoising due to low signal-to-noise ratios. Patch2Self (P2S) employing self-supervised techniques and regression on a Casorati matrix effectively denoises dMRI images and has become the new de-facto standard in this field. P2S however is resource intensive both in terms of running time and memory usage as it uses all voxels (n) from all-but-one held-in volumes (d-1) to learn a linear mapping Phi : \mathbb R ^ n x(d-1) \mapsto \mathbb R ^ n for denoising the held-out volume. The increasing size and dimensionality of higher resolution dMRI acquisitions can make P2S infeasible for large-scale analyses. This work exploits the redundancy imposed by P2S to alleviate its performance issues and inspect regions that influence the noise disproportionately. Specifically this study makes a three-fold contribution: (1) We present Patch2Self2 (P2S2) a method that uses matrix sketching to perform self-supervised denoising. By solving a sub-problem on a smaller sub-space so called coreset we show how P2S2 can yield a significant speedup in training time while using less memory. (2) We present a theoretical analysis of P2S2 focusing on determining the optimal sketch size through rank estimation a key step in achieving a balance between denoising accuracy and computational efficiency. (3) We show how the so-called statistical leverage scores can be used to interpret the denoising of dMRI data a process that was traditionally treated as a black-box. Experimental results on both simulated and real data affirm that P2S2 maintains denoising quality while significantly enhancing speed and memory efficiency achieved by training on a reduced data subset.

As large-scale foundation models become publicly available for different domains efficiently adapting them to individual downstream applications and additional data modalities has turned into a central challenge. For example foundation models for geospatial and satellite remote sensing applications are commonly trained on large optical RGB or multi-spectral datasets although data from a wide variety of heterogeneous sensors are available in the remote sensing domain. This leads to significant discrepancies between pre-training and downstream target data distributions for many important applications. Fine-tuning large foundation models to bridge that gap incurs high computational cost and can be infeasible when target datasets are small. In this paper we address the question of how large pre-trained foundational transformer models can be efficiently adapted to downstream remote sensing tasks involving different data modalities or limited dataset size. We present a self-supervised adaptation method that boosts downstream linear evaluation accuracy of different foundation models by 4-6% (absolute) across 8 remote sensing datasets while outperforming full fine-tuning when training only 1-2% of the model parameters. Our method significantly improves label efficiency and increases few-shot accuracy by 6-10% on different datasets.

Parameter-efficient fine-tuning (PEFT) is an effective methodology to unleash the potential of large foundation models in novel scenarios with limited training data. In the computer vision community PEFT has shown effectiveness in image classification but little research has studied its ability for image segmentation. Fine-tuning segmentation models usually require a heavier adjustment of parameters to align the proper projection directions in the parameter space for new scenarios. This raises a challenge to existing PEFT algorithms as they often inject a limited number of individual parameters into each block which prevents substantial adjustment of the projection direction of the parameter space due to the limitation of Hidden Markov Chain along blocks. In this paper we equip PEFT with a cross-block orchestration mechanism to enable the adaptation of the Segment Anything Model (SAM) to various downstream scenarios. We introduce a novel inter-block communication module which integrates a learnable relation matrix to facilitate communication among different coefficient sets of each PEFT block's parameter space. Moreover we propose an intra-block enhancement module which introduces a linear projection head whose weights are generated from a hyper-complex layer further enhancing the impact of the adjustment of projection directions on the entire parameter space. Extensive experiments on diverse benchmarks demonstrate that our proposed approach consistently improves the segmentation performance significantly on novel scenarios with only around 1K additional parameters.

Applying a pre-trained large model to downstream tasks is prohibitive under resource-constrained conditions. Recent dominant approaches for addressing efficiency issues involve adding a few learnable parameters to the fixed backbone model. This strategy however leads to more challenges in loading large models for downstream fine-tuning with limited resources. In this paper we propose a novel method for increasing the parameter efficiency of pre-trained models by introducing an intermediate pre-training stage. To this end we first employ low-rank approximation to compress the original large model and then devise a feature distillation module and a weight perturbation regularization module. These modules are specifically designed to enhance the low-rank model. In particular we update only the low-rank model while freezing the backbone parameters during pre-training. This allows for direct and efficient utilization of the low-rank model for downstream fine-tuning tasks. The proposed method achieves both efficiencies in terms of required parameters and computation time while maintaining comparable results with minimal modifications to the backbone architecture. Specifically when applied to three vision-only and one vision-language Transformer models our approach often demonstrates a merely 0.6-point decrease in performance while reducing the original parameter size by 1/3 to 2/3.

Customization techniques for text-to-image models have paved the way for a wide range of previously unattainable applications enabling the generation of specific concepts across diverse contexts and styles. While existing methods facilitate high-fidelity customization for individual concepts or a limited pre-defined set of them they fall short of achieving scalability where a single model can seamlessly render countless concepts. In this paper we address a new problem called Modular Customization with the goal of efficiently merging customized models that were fine-tuned independently for individual concepts. This allows the merged model to jointly synthesize concepts in one image without compromising fidelity or incurring any additional computational costs. To address this problem we introduce Orthogonal Adaptation a method designed to encourage the customized models which do not have access to each other during fine-tuning to have orthogonal residual weights. This ensures that during inference time the customized models can be summed with minimal interference. Our proposed method is both simple and versatile applicable to nearly all optimizable weights in the model architecture. Through an extensive set of quantitative and qualitative evaluations our method consistently outperforms relevant baselines in terms of efficiency and identity preservation demonstrating a significant leap toward scalable customization of diffusion models.

Visual object tracking aims to localize the target object of each frame based on its initial appearance in the first frame. Depending on the input modility tracking tasks can be divided into RGB tracking and RGB+X (e.g. RGB+N and RGB+D) tracking. Despite the different input modalities the core aspect of tracking is the temporal matching. Based on this common ground we present a general framework to unify various tracking tasks termed as OneTracker. OneTracker first performs a large-scale pre-training on a RGB tracker called Foundation Tracker. This pretraining phase equips the Foundation Tracker with a stable ability to estimate the location of the target object. Then we regard other modality information as prompt and build Prompt Tracker upon Foundation Tracker. Through freezing the Foundation Tracker and only adjusting some additional trainable parameters Prompt Tracker inhibits the strong localization ability from Foundation Tracker and achieves parameter-efficient finetuning on downstream RGB+X tracking tasks. To evaluate the effectiveness of our general framework OneTracker which is consisted of Foundation Tracker and Prompt Tracker we conduct extensive experiments on 6 popular tracking tasks across 11 benchmarks and our OneTracker outperforms other models and achieves state-of-the-art performance.